Publications 2023

Detection and characterisation of hidden damages in layered composites like Fibre laminates, e.g., Fibre Metal Laminates (FML), is still a challenge. Commonly, Guided Ultrasonic Waves (GUW) or X-ray imaging are used to detect hidden damages. X-ray imaging can be divided into two-dimensional transmission or reflection and three-dimensional tomography imaging using reconstruction algorithms to compute a three-dimensional view from slice images. Damages or defects can be classified roughly in layer delaminations, extended cracks, micro cracks (fibres and solid material layer), deformations, and impurities during manufacturing. Detection of such kind of damages and defects by visual inspection is a challenge, even using 3D CT data, and moreover using single 2D transmission images. For damage characterisation, micro-focus CT X-ray scanner are used, providing a high resolution below 100 μm, but with the disadvantage of high scanning times (up to several hours) [CHA22]. Anomaly detectors based on advanced data-driven Machine Learning methods can be used to mark Regions-of-Interest (ROI) in images automatically (feature selection process). ROI feature extraction is the first stage of an automated damage diagnostic system providing damage detection, classification, and localisation. But data-driven methods require typically a sufficient large set of training examples (with respect to diversity and generality), which cannot be provided commonly in engineering and damage diagnostics (e.g., an impact damage can only be "created" one time and is not reversible). In this work, the challenges, limits, and detection accuracy of automated ROI damage feature detection from low-quality and low-resolution 2D X-ray image data using data-driven anomaly detectors are investigated and evaluated. In addition to experimental data, X-ray simulation is used to create an augmented training and test data set. Simulated and experimental X-ray data are compared. The simulation is carried out with the gvirtualxray [GVX23;VID21] software It is based on the Beer-Lambert law to compute the absorption of light (i.e. photons) by 3D objects (here polygon meshes). Additionally, X-ray ray-tracing by the the x-ray projection simulator [DAC23] software is used for comparison. Suitable data-driven anomaly detectors estimating and marking the ROI candidates of damage areas are Convolutional Neural Networks trained supervised (i.e., using manually feature labelled data), either used as a pixel-based feature classifier (Point-Net) or as a region-based proposal network (Region-based CNN, R-CNN, Fast R-CNN, Region-proposal networks) [KHA18].The generated knowledge and the image data collected would further accelerate the development in the field of autonomous SHM of the composite structures which would further reduce the safety risks and total time associated with structural integrity assessment. References: [KHA18] S. Khan, H. Rahmani, S. A. A. Shah, and M. Bennamoun, A Guide to Convolutional Neural Networks for Computer Vision. Morgan & Claypool Publishers, 2018; [GVX23] gvirtualxray, https://gvirtualxray.fpvidal.net, accessed on-line on 24.1.2023; [VID21] F. P. Vidal, Introduction to X-ray simulation on GPU using gVirtualXRay, In Workshop on Image-Based Simulation for Industry 2021 (IBSim-4i 2020), London, UK, October, 2021; [DAC23] X-Ray Projection Simulator based on Raytracing, https://git.scc.kit.edu/dach/raytracingx-rayprojectionsimulator, access on-line on 24.1.2023; [CHA22] C. Shah, S. Bosse, and A. von Hehl, “Taxonomy of Damage Patterns in Composite Materials, Measuring Signals, and Methods for Automated Damage Diagnostics,” Materials, vol. 15, no. 13, p. 4645, Jul. 2022, doi: 10.3390/ma15134645.

Optimization of material distribution with the aim of maximizing stiffness is a common problem in engineering design aimed at structures offering low weight and/or limited design space, and several solutions are known [1]. In parallel, manufacturing techniques are being developed which allow realization of arbitrary material distribution in multi-material components: Typically, these focus on bi-material structures and are based either in casting, using a compound casting approach in which one material is introduced as insert around which the other is cast [2], or in additive manufacturing, where multi-material processes are increasingly being realized for several material classes, including polymers and metals [3,4]. The present study considers the Multi Phase Topology Optimization (MPTO) approach originially suggested by Burblies and Busse and further studied by Mounchili at al. [5,6]. This technique is based on iterative linear-elastic FEM simulations and the evaluation of strain energy data on model as well as element level and builds on the fact that association of high stiffness material properties with elements experiencing high levels of strain energy will serve to minimize total strain energy, and thus decrease displacement under a given load. The analysis covers (a) the realization of the MPTO approach based on different algorithms for adapting the material distribution and (b) its capability of identifying best combinations of low weight and high stiffness for a given load case and design space subject to variation of material volume fractions. Fundamental investigations on a completely stochastic and a genetic algorithm, with and without added integration of a physics-based sorting approach, are performed on a simple load case and limited model size to support fast optimization runs, thus allowing scrutiny of the scatter of results. The weight optimization problem is addressed using the same model of an asymmetric three-point bending setup incorporating equal fractions of three materials. A final validation is performed on a more complex model and load case with a the number of degrees of freedom increased by two orders of magnitude. The presentation closes with an outlook on further development paths, which include, on the computational side, pre-filtering of configurations to reduce the number of FEM simulations e.g. via integration of a machine learning-based predictor function, and on the materials engineering side the consideration of material interface characteristics as well as extensions towards incorporating aspects of plasticity. References [1] H. Z. Yu, S. R. Cross, C. A. Schuh Journal of Materials Science, 2017, 52, 4288-4298. [2] D. Schittenhelm, A. Burblies, M. Busse Forschung im Ingenieurwesen, 2018, 82, 131-147. [3] Y. Zheng, W. Zhang, D. M. Baca Lopez, R. Ahmad Polymers, 2021, 13, 1957. [4] A. Bandyopadhyay, B. Heer Mat. Sci. Eng. R: Reports, 2018, 129, 1-16. [5] A. Burblies, M. Busse Proceedings of the Multiscale & Functionally Graded Materials Conference (FGM), October 15th-18th, 2006, Honolulu (Hawaii, USA) [6] A.P. Mounchili, S. Bosse, D. Lehmhus, A. Struss MATEC Web of Conferences, 2021, 349, 03001.

In the case of medical products, in particular implants, high demands are placed on the materials and their properties. With additively manufactured implants made of Ti6Al4V, porosity develops due to the process. In terms of process technology, this is kept to a level of less than one percent and is often almost completely closed by a connected thermo-mechanical after treatment, HIP for short (hot isostatic pressing). Thus, the end products have almost no recognizable porosity anymore. , There may be significant differences in the mechanical properties. In this case, especially with the dynamic loading of the components. Since a failure of implants causes a highly sensitive situation, the relationship between the process parameters in the manufacturing process and the mechanical properties must be investigated. Of particular interest is the influence of the final properties by the HIP. Using models of machine learning and image analysis, differences in the microstructure such as Melt trace size and shape, Grain size, Phase components, alpha and Beta, Phase characteristics (shape, size and position), Grain and phase orientation, as well as Pores [1] can be recognized, described and associated with the process parameters and mechanical properties. Here, samples in the as-built as well as in the hyped state are to be examined. Challenges exist in particular in the differentiation and identification of features in the very fine microstructure as well as the relatively small number of laboratory tests due to the experimental and preparation effort. An experimental design will be developed in cooperation between the data sciences and the materials sciences, which should lead to a continuous refinement of the models through an iterative procedure. Proposed data science and analysis methods and algorithms: Density-based clustering, CNN / Region-based CNN, F-RSN, Residual Neural Network, Autoencoder-based anomaly detectors, Numerical approaches, image transformations, clustering analysis. Formally, there is a model predictor function f(x):x -> y, with input x as measuring data from a set of experiments and specimens, and output y characterizing features of the manufactured material. The target features have relevant impact on the lifecycle and robustness of medical implants, which have to be identified, too. The input data is heterogeneous, but often it consists of images, therefore image analysis and object detection (with ROI) are fundamental pre-processing steps [2]. There is a large set of already available object detectors, e.g., coco-ssd. A major challenge in object detection is the complexity of ROi proposals with respect to model complexity and computational complexity. Using pure data-driven models, e.g.,coco-ssd, trained for environmental scene recognition, the ROI proposal and object detection in measuring images, e.g., from pore micrographs, will result in bad coverage and accuracy. For this reason, we will apply model-assisted object detectors, e.g., to find critical material pores, to identify cracks, and different material regions. References: [1] M. L. Altmann, S. Bosse, C. Werner, R. Fechte-Heinen, and A. Toenjes, Programmable Density of Laser Additive Manufactured Parts by Considering an Inverse Problem, Materials, vol. 15, no. 20, p. 7090, Oct. 2022, doi: 10.3390/ma15207090; [2] C. Shah, C., S. Bosse, A. von Hehl, Materials 2022, 15, 4645. https://doi.org/10.3390/ma15134645